20 Prompt Engineering & Structured Output Practice Questions & Answers

1. 1. An automated code-review pipeline reports that 40% of its "inaccurate comment" findings are dismissed by developers. The prompt currently reads: "Check that code comments are accurate. Be conservative and only report high-confidence issues." Which change most effectively reduces the false-positive rate?

   • A.Replace the vague instruction with an explicit categorical criterion: report a comment only when the behavior it claims contradicts what the code actually does, and skip merely outdated wording or style✓ Answer
   • B.Strengthen the hedge to "be extremely conservative and report only issues you are absolutely certain about"
   • C.Train a separate classifier on historically dismissed findings and use it to pre-filter comment issues
   • D.Have the model emit a 0–1 confidence score with each finding and discard anything below 0.9

   Generic hedges such as "be conservative" or "only high-confidence issues" give the model no decision boundary — they change tone, not the rule the model applies, so precision does not improve. A specific categorical criterion names the exact condition that makes something reportable, which is what actually separates true positives from noise. Self-reported confidence scores are uncalibrated and unreliable as a filter, and a bespoke classifier is heavy over-engineering for a problem solvable in the prompt.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.1 (explicit criteria over vague instructions)Report a problem with this question

2. 2. A review prompt instructs the model to "report anything that looks problematic." Output mixes genuine correctness bugs with naming preferences and formatting nits, and the team has started ignoring the report entirely. What is the most effective change?

   • A.Append "focus only on what really matters" to the end of the prompt
   • B.Run the review three times and keep only the findings that appear in at least two runs
   • C.Add a reasoning step so the model deliberates before producing each finding
   • D.Enumerate in the prompt which categories are in scope (correctness defects, security flaws, data-loss risks) and which are explicitly out of scope (formatting, naming, local stylistic conventions)✓ Answer

   "Problematic" and "what really matters" are undefined terms, so the model supplies its own scope and the report becomes unusable. Explicitly enumerating in-scope and out-of-scope categories gives a deterministic filter the model can apply consistently. Consensus voting across runs is a known anti-pattern here: real bugs that are caught only intermittently get suppressed, and a reasoning step does not fix a scope that was never defined.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.1 (defining which issues to report versus skip)Report a problem with this question

3. 3. A team has rewritten an extraction prompt six times over three weeks. Each rewrite is based on a developer's impression of recent output, and accuracy has not measurably improved. What is the most effective way to make prompt iteration productive?

   • A.Ship each new version straight to production and treat the volume of user complaints as the evaluation signal
   • B.Add a standing instruction telling the model to avoid the mistakes it made in previous versions
   • C.Rewrite the prompt from scratch each cycle, adding more instructions than the previous version contained
   • D.Assemble a fixed set of real failing cases with known correct answers, change one element of the prompt at a time, and re-score every version against that same set✓ Answer

   Prompt changes are only improvements if they can be measured against a stable reference; without a held-out set of real failures and known-correct answers, each rewrite trades one failure mode for another invisibly. Changing one element at a time preserves attribution, so the team learns which change caused which effect. The model has no memory of prior prompt versions, so instructing it to "avoid previous mistakes" references context it does not possess.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.5 (prompt refinement on a sample set before scaling) and 4.1Report a problem with this question

4. 4. A document-extraction prompt already contains three paragraphs of detailed formatting instructions, yet the output shape still varies noticeably from run to run on similar documents. Which technique is most likely to produce consistent output?

   • A.Add the sentence "you must always follow the format exactly, without exception"
   • B.Add a small set of worked examples showing a representative input and the exact output it should produce✓ Answer
   • C.Ask the model to reason at length about formatting before it produces the extraction
   • D.Repeat the formatting instructions in both the system prompt and the user turn so the model sees them twice

   When detailed prose instructions alone still yield inconsistent output, few-shot examples are the single most effective remedy: a demonstrated input–output pair conveys structure the model can copy directly, whereas prose must first be interpreted into a structure. Repeating or emphasizing the same instruction adds tokens without adding information, and reasoning about formatting does not supply the missing target shape.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.2 (few-shot examples as the most effective technique for consistent output)Report a problem with this question

5. 5. An engineer proposes pasting 30 example input/output pairs into an ambiguous-case classification prompt, arguing that more examples always help. Cost and latency rise sharply while accuracy gains flatten after the first few. What guidance should the architect give?

   • A.Use as many examples as the request will physically accommodate, since example count is the dominant quality factor
   • B.Reduce to a single canonical example, because one clear demonstration is always sufficient
   • C.Use roughly 2–4 targeted examples, each chosen to illustrate a distinct decision boundary rather than to repeat one already covered✓ Answer
   • D.Keep all 30 but shorten each one so the total length stays manageable

   The value of a few-shot example comes from the distinct judgment it demonstrates, not from its presence in a tally, so a small targeted set of roughly 2–4 examples that each mark a different decision boundary captures nearly all the available benefit. Beyond that the examples become redundant and consume budget without teaching anything new. A single example is usually too narrow to convey where the boundary lies for ambiguous cases.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.2 (creating 2–4 targeted few-shot examples for ambiguous scenarios)Report a problem with this question

6. 6. An invoice-extraction prompt contains six few-shot examples, all drawn from the same clean vendor template. Accuracy is 96% on that template but drops to 61% on scanned invoices, multi-currency invoices, and invoices where totals appear in a footer. What is the best fix?

   • A.Replace the six near-identical examples with a few examples that each cover a different structural variant, including one ambiguous case that shows why a particular field was chosen✓ Answer
   • B.Add the instruction "do not fail on unusual invoice layouts"
   • C.Ask the model to state its confidence for each field so low-confidence layouts can be skipped
   • D.Add twenty more examples from the same vendor template to reinforce the pattern

   Near-identical examples teach pattern matching against one layout rather than the underlying judgment, which is exactly why performance collapses off-template. Examples that span the real range of structural variants — and that show the reasoning for choosing one field over a plausible alternative — let the model generalize to layouts it has never seen. Adding more of the same examples deepens the overfit, and a prohibition supplies no new information about how to read a different layout.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.2 (few-shot examples for varied document structures; generalizing judgment to novel patterns)Report a problem with this question

7. 7. A summarization prompt describes the desired output in prose: "Produce a short heading, then three to five bullets, then a single-line recommendation." In practice the model sometimes merges the bullets into a paragraph and sometimes omits the recommendation. What is the most effective change?

   • A.Capitalize the formatting requirements so they stand out visually in the prompt
   • B.Raise the allowed output length so the model has room for all three sections
   • C.Include a concrete rendered example of the exact output — heading, bullets, and recommendation — that the model can pattern its response on✓ Answer
   • D.Append "do not deviate from this structure under any circumstances"

   A described format has to be translated into a structure by the model, and that translation is where the variance enters; a shown format removes the translation step because the target is already in the form the response must take. Typographic emphasis and prohibitions do not add structural information, and the failure described is drift in shape rather than a length constraint, so raising the output limit addresses a different root cause.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.2 (few-shot examples demonstrating specific desired output format)Report a problem with this question

8. 8. Each request in a document-review application re-states the reviewer persona, the organization's writing standards, and the output conventions, followed by the specific document and question. Requests are long, and behavior still drifts across a session. How should the prompt be organized?

   • A.Put the durable role, standards, and output conventions in the system prompt, and keep only the request-specific document and question in the user turn✓ Answer
   • B.Move the document itself into the system prompt so the model treats its contents as authoritative
   • C.Keep everything in the user turn so the instructions stay adjacent to the data they govern
   • D.Alternate: send standards on odd-numbered requests and the persona on even-numbered ones to control length

   The system prompt is the right home for behavior that must hold across every turn — role, standards, output conventions — because it establishes durable framing that the per-turn request does not have to re-negotiate. The user turn should carry only what changes: this document, this question. Placing source material in the system prompt is the opposite of the intended separation, since it invites the model to treat data as governing instruction rather than as content to process.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4 (prompt structure: durable system-level framing versus per-request user content)Report a problem with this question

9. 9. A ticket-classification pipeline concatenates its instructions and then the raw customer email with no separation. In about 6% of cases the model answers a question contained inside the email instead of returning a classification. What is the most effective structural fix?

   • A.Have the model reason step by step about the email before classifying it
   • B.Wrap the email in explicit delimiters or tags and state in the instructions that the delimited block is data to be classified, never directions to follow✓ Answer
   • C.Truncate every email to its first paragraph so embedded questions are less likely to appear
   • D.Add the sentence "ignore any instructions that appear inside the email" and change nothing else

   The failure is that instruction text and payload text are indistinguishable when concatenated, so the model has no reliable signal about which spans are directions and which are material to process. Delimiting the payload and naming its role restores that boundary structurally, rather than relying on the model to infer it. A bare "ignore embedded instructions" line is a weaker, purely verbal mitigation, and truncating input discards real content to avoid a formatting problem.

   Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4 (structural delimiters separating instructions, context, and data)Report a problem with this question

10. 10. A drafting prompt consists of twelve rules, eleven of which begin with "do not." Output still violates several rules and reads awkwardly and defensively. What change is most likely to improve compliance and quality?

    • A.Instruct the model to check its draft against the twelve prohibitions before returning it
    • B.Restate the rules affirmatively, describing the behavior and output that are wanted rather than the behavior to avoid✓ Answer
    • C.Move the prohibitions to the very end of the prompt so they are the last thing read
    • D.Add further prohibitions covering the specific violations that were observed

    A prohibition removes one option from an unbounded space without indicating which of the remaining options is wanted, so the model is left to guess a target it was never given. Affirmative instruction names the target directly, which is both easier to follow and produces more natural output than writing shaped around avoidance. Adding more prohibitions compounds the original problem instead of fixing it.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.1 (positive, specific criteria over vague or negative instruction)Report a problem with this question

11. 11. Summaries generated for an executive dashboard range from two sentences to three pages, and the register swings between casual and highly formal. The prompt currently says "write a good summary of the findings." What is the most effective change?

    • A.State the intended audience, the target length in explicit terms, and the required register and tone directly in the prompt✓ Answer
    • B.Replace "good" with "concise and professional"
    • C.Set a hard maximum output length so responses cannot run long
    • D.Ask the model to draft the summary and then revise it until it is satisfied with the style

    Length, tone, and register have no correct default — if the prompt does not specify them, the model selects plausible values that differ from run to run, which is exactly the variance observed. Naming the audience, the length target, and the register converts three unstated variables into stated requirements. A hard output cap addresses a different root cause: it truncates a long response rather than shaping a right-sized one, and "concise and professional" is still an undefined term.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4 (specifying audience, format, length, and success criteria explicitly)Report a problem with this question

12. 12. A benefits-eligibility assistant must combine income thresholds, household size, and residency duration to reach a determination. It states its conclusion immediately and is wrong in roughly 15% of multi-condition cases, though each individual rule is stated correctly in the prompt. What change is most likely to help?

    • A.Have the model work through each condition explicitly before stating the determination, and place the final determination in a clearly designated part of the output✓ Answer
    • B.Add a prohibition telling the model never to guess when it is unsure
    • C.Add more examples that show only the final determination for each case
    • D.Shorten the prompt so the rules are closer together and easier to hold in view

    When a correct answer depends on composing several intermediate results, asking for the conclusion first forces the model to commit before the composition has been performed. Making the intermediate steps explicit lets each condition be resolved and checked in turn, and designating where the final answer goes keeps the output machine-readable despite the added reasoning. Examples showing only final answers demonstrate the destination without the path, so they do not repair multi-step composition.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4 (reasoning before answering for multi-step analysis)Report a problem with this question

13. 13. A team adds "think step by step before answering" to a high-volume prompt that assigns each incoming message to one of four fixed categories. Accuracy is unchanged, per-request latency roughly doubles, and cost rises. What should the architect recommend?

    • A.Remove the reasoning step for this task, reserving explicit reasoning for prompts that require multi-step analysis rather than a single categorical judgment✓ Answer
    • B.Keep the reasoning step and add examples of well-reasoned classifications so the reasoning becomes useful
    • C.Keep the reasoning step and move the workload to overnight processing to absorb the latency
    • D.Keep the reasoning step but instruct the model to limit it to one sentence

    Explicit reasoning pays off when the answer must be composed from intermediate results; a single assignment to one of four fixed categories has no intermediate results to compose, so the reasoning adds tokens, latency, and cost while producing no accuracy gain — which the measurements confirm. The correct response is to remove a technique that is not earning its cost, not to trim it, hide its latency behind asynchronous processing, or invest further in it.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4 (matching technique to task: reasoning for multi-step analysis, not simple classification)Report a problem with this question

14. 14. A prompt instructs the model to "flag any passage that violates our editorial style guide and format the result the way we normally do." Neither the style guide nor a sample of the usual format is included in the request. Output is inconsistent and often flags the wrong things. What is the root cause?

    • A.The model needs an explicit reasoning step before it can apply editorial judgment
    • B.The task is too complex for a single request and must be decomposed regardless of what context is supplied
    • C.The instruction depends on shared context the model does not have, so the model substitutes its own generic notion of style and format✓ Answer
    • D.The prompt lacks a prohibition against flagging passages that are stylistically acceptable

    "Our style guide" and "the way we normally do it" are references to information that exists only inside the organization; the model cannot retrieve them, so it fills the gap with a generic default that differs run to run. Instructions that assume shared context fail for this reason, and the fix is to include the actual criteria and a sample of the target format in the prompt. Decomposition, reasoning steps, and prohibitions all presuppose criteria that were never supplied.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.1 (explicit criteria; instructions that assume unavailable context fail)Report a problem with this question

15. 15. An extraction service ends its prompt with "Respond only with JSON. Do not include any other text." The downstream parser still fails on roughly 7% of responses because of a leading sentence of prose or a surrounding code fence. What is the most reliable fix?

    • A.Add three examples of correctly formatted JSON-only responses to the prompt
    • B.Require the output to conform to a declared schema through the structured tool-call mechanism, so the model returns schema-conformant fields rather than free-form text✓ Answer
    • C.Add a stronger admonition such as "any extra text will break production systems"
    • D.Post-process every response with a regular expression that strips prose and code fences before parsing

    Asking politely for JSON is a probabilistic constraint on free-form text generation, so a residual failure rate is expected no matter how the request is worded. Requiring conformance to a declared schema is a structural constraint: the output is produced as schema-conformant fields, so the class of prose-preamble and code-fence failures disappears by construction rather than by persuasion. Regex stripping patches the symptom and will itself fail on unanticipated wrappers, and examples reduce but do not eliminate the drift.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.3 (tool use with JSON schemas as the most reliable approach for guaranteed schema-compliant output)Report a problem with this question

16. 16. After moving an invoice pipeline to schema-constrained structured output, parse errors drop to zero. However, in 12% of invoices the extracted line items do not sum to the extracted total, and shipping charges sometimes land in the tax field. What does this indicate, and what is the appropriate response?

    • A.A schema constrains the shape of the output but not its truthfulness, so add a computed total alongside the stated total, validate the relationship in code, and route mismatches to review✓ Answer
    • B.The extraction should be retried with the identical prompt until the totals happen to agree
    • C.The parser is misconfigured, since a conforming response cannot contain incorrect values
    • D.The schema is not strict enough; marking more fields as required will resolve both problems

    Schema enforcement eliminates syntax errors — malformed structure, missing braces, wrong types — but it cannot detect semantic errors such as values that do not sum or values placed in the wrong field, because those responses are perfectly well-formed. Surfacing a calculated total next to the stated total turns an invisible semantic error into a checkable programmatic assertion. Adding required fields tightens shape, not truth, and blind retries do not change the underlying reasoning.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statements 4.3 and 4.4 (schemas eliminate syntax errors but not semantic errors; calculated_total alongside stated_total)Report a problem with this question

17. 17. A contract-extraction schema marks contract_end_date as a required field. About 30% of the contracts in the corpus are open-ended and state no end date, and for those the model returns plausible but invented dates. What is the correct fix?

    • A.Add a second verification pass in which the model re-reads the contract and checks the date it produced
    • B.Add the instruction "never invent a date that is not present in the contract" to the prompt
    • C.Make the field optional or explicitly nullable so that "no end date present" is a valid, representable answer✓ Answer
    • D.Retry the extraction with the validation error appended until a correct date is returned

    A required field with no corresponding source data creates structural pressure to fabricate: the schema declares that a value must exist, so the model produces one. The fix is schema redesign — making absence representable — not additional prompt language or an extra checking layer, both of which leave the contradictory requirement in place. Retrying is futile here because the information is genuinely absent from the source, and no number of attempts can extract what is not there.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.3 (designing fields as optional/nullable to prevent fabrication) and 4.4 (limits of retry)Report a problem with this question

18. 18. A classification schema offers six fixed document-type values. Reviewers find that genuinely ambiguous documents and legitimate new document types are being forced into the closest existing value, and the errors are invisible downstream. What schema change best addresses this?

    • A.Expand the enumeration to forty values so that nearly every document has an exact match
    • B.Remove the enumeration entirely and let the model return a free-text document type
    • C.Add a prompt instruction telling the model not to classify documents it is unsure about
    • D.Add an "unclear" value for ambiguous cases and an "other" value paired with a free-text detail field for types outside the current taxonomy✓ Answer

    A closed enumeration with no escape hatch makes a wrong answer the only expressible answer, so ambiguity and novelty are silently converted into confident misclassification. Adding "unclear" and "other" plus a detail string makes those two states representable, which both stops the forced misfit and surfaces the cases that need human attention or taxonomy expansion. Dropping the enumeration sacrifices the consistency the schema exists to provide, and a prompt instruction cannot express a value the schema disallows.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.3 (enum values such as "unclear" and "other" plus detail fields)Report a problem with this question

19. 19. A batch report generator produces long structured outputs. In a small fraction of runs the output ends mid-record, and the response indicates that generation halted because the output-length limit was reached rather than because the model finished. What is the correct handling?

    • A.Detect the length-limit signal, discard the partial output rather than parsing it, and reduce the scope of each request by splitting the input into smaller units or requesting a more compact output shape✓ Answer
    • B.Parse whatever was returned and treat the missing records as absent from the source
    • C.Resubmit the identical request unchanged, since the same prompt may fit next time
    • D.Interpret the truncation as a refusal and rewrite the instructions to be more permissive

    Output that stops at a length limit is incomplete rather than wrong, and the distinction matters: parsing it silently converts a truncation into fabricated data loss, because absent records look identical to records the source never contained. The response's own halt signal is the reliable way to tell the two apart. Since the cause is that the requested output exceeds the available room, the remedy is to shrink what each request must emit — smaller input units or a more compact shape — not an identical retry, which faces the same limit.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.4 (handling non-conforming output; distinguishing structural output failures from absent source information)Report a problem with this question

20. 20. A single prompt asks the model to read a 90-page solicitation, extract every requirement, assess feasibility against internal capabilities, draft a narrative response, and produce a pricing table. Results are shallow, later sections are weaker than earlier ones, and quality varies unpredictably. What is the most effective change?

    • A.Add a reasoning step instructing the model to think carefully about each of the five tasks
    • B.Append "be thorough and do not skip any section of the solicitation"
    • C.Decompose the work into a sequence of separately scoped prompts, each with one objective and a defined output that feeds the next stage✓ Answer
    • D.Add few-shot examples of complete finished proposals so the model knows what the end product looks like

    Five distinct objectives competing inside one request divide attention across all of them, which is why depth falls off and later sections degrade — no single objective ever receives focused treatment. Splitting the work into scoped prompts gives each stage a single goal and a defined output that the next stage consumes, so each step can be evaluated and corrected independently. Reasoning steps, exhortations, and finished-product examples all leave the overloaded single request intact.

    Source: Anthropic Claude Certified Architect – Foundations, Anthropic Claude certification (Architect – Foundations), Domain 4, Task Statement 4.6 (decomposing oversized work into focused passes to avoid attention dilution)Report a problem with this question

Practice questions based on the official Claude Certified Architect – Foundations (CCAR-F) exam guide and Anthropic's public documentation. This is an independent study tool, not affiliated with or endorsed by Anthropic, and does not grant certification. The real exam is 60 questions, 120 minutes, passing at a scaled 720/1000, delivered via Pearson VUE ($125). Official certification page →